smartbear0902/sn92-generalist-s1

REAX 1 generalist fine-tune (SN92)

A fine-tune of decisionmodels/reax-1 (revision ca2a0cc54a537c1f5cf24e5b09387cfb357b6518) for the Bittensor subnet 92 (REAX) generalist decision-model track. Same architecture, joint readout head and runtime as the parent: run it exactly like the parent.

Recipe

A weighted merge ("model soup") of 3 fine-tune checkpoints: the weighted sum of their LoRA deltas and of their readout-head deltas, added to the parent:

Fine-tune checkpoint	Merge weight
run A (data mix A, 19,850 items), final checkpoint after 877 steps	0.34
run C (data mix A plus extra arithmetic, code, rules and registry items, 23,654 items), step 300 of 1,096	0.33
run C, step 450 of 1,096	0.33

Each run: LoRA rank 16, alpha 32, on the 496 projection matrices of the language model (full-attention q/k/v/o, linear-attention in_proj_qkv/z/a/b and out_proj, MLP gate/up/down) with the readout head trained jointly; objective: cross-entropy of the answer code at the runtime temperature T = 2 plus a KL term (weight 0.5) to the unmodified parent; learning rate 1e-4 (readout 1e-5); one epoch, 32 question rows per step.

Training data: procedurally generated decision items from our own generators (arithmetic word problems, multi-step logic, causal yes/no, code output, chords from notes, tool retrieval / next API / when-to-call / tool-call decisions, tariff and policy rules; CC0) and public datasets under their licences: Aegis 2.0 (CC BY 4.0), MASSIVE (CC BY 4.0), IBM ArgQ (CC BY-SA 3.0), MultiNLI (OANC / CC BY-SA 3.0 parts), arXiv metadata (CC0), Standard Ebooks (CC0), openFDA and NVD (public domain), Kubernetes enhancement proposals (Apache-2.0) and Debian package metadata.

Data provenance

• No item of the REAX eval-v5 public dev set (eval-v5/devset-v1), of any validator draw or of any other SN92 private pool was used for training, for checkpoint selection inside a run, or to build the generators' item lists. The public dev set was used only to measure finished checkpoints, as the FAQ intends.
• The procedural items come from our own generators (code available to REAX on request); the public datasets are used under their licences as listed above. No model outputs of other SN92 entries were used as training targets.
• A larger gain on public-style items than on the private draws is expected for any targeted fine-tune; we are happy to answer questions from the REAX review in #92.

Credit

REAX 1 is by REAX / decisionmodels.io (Apache-2.0), a derivative of Perplexity AI's pplx-decider-v1.2, itself derived from Qwen/Qwen3.8-27B. See NOTICE and MODIFICATIONS.md.